Floating seismometers (‘MERMAIDs’) operating in the noisy environment of the world’s oceans pose a challenge for picking the time of earthquake first arrivals. We report on an experiment to estimate the errors in picked arrivals from 49 MERMAIDS operating in the South Pacific, using two independent strategies. For 15 events, the same arrivals were redundandly picked by several analysts, allowing for a direct estimate of error distributions. Standard errors in times from MERMAID seismograms vary from 0.2 s for close events at mantle depths in the Kermadec subduction zone to more than 2 s for crustal events at large epicentral distance. In a second experiment we analysed the a posteriori misfits after tomographically inverting all events. The residual traveltime misfit is consistent with the error estimates from the first experiment, but also shows inconsistencies with arrival times from the ISC-EHB and NEIC catalogues, which we attribute to errors in the published hypocentres and/or origin times.